Part 3: Making Warehouses Smarter

Making Warehouses Smarter

Adaptive Algorithm for Order Batching & Sequencing applicable to all pick workflows


Manual order picking is one of the most core activities in an e-commerce warehouse/fulfillment center. A customer order in a typical e-commerce scenario consists of multiple order lines belonging to different inventory items, each with a specific quantity in demand. The pickers must travel within/across the zones in the warehouse to complete the picks required for each order. However, the warehouse floor usually spans across acres, making the manual picking process a very time-consuming and costly operation constituting towards more than 50% of the total warehouse operations cost.

The total worldwide e-commerce sales have grown more than 325% in the last 8 years, and it is further expected to grow by 45% in the coming 3 to 4 years. The average number of order lines (items) in an order is also continuously increasing with an increase of 1 item per order on average in the last 3-4 years. Owing to these current trends, warehouses are observing a constant surge in the number of orders that they must fulfill in a day, prompting these warehouses to decrease their order fulfillment time as much as possible. The warehouse floor area limits the number of pickers which can be employed in a shift and adding more employees can reduce efficiency and further increase the order fulfillment cost. Under such rapidly changing circumstances, optimization of available resources becomes very essential to effectively utilize the labor, reduce the order fulfillment time, and enhance the picking throughput. 

Traditionally, Batch-Picking is implemented in most of the warehouses to optimize the manual picking process. In Batch-Picking, a pool of 100-200 orders (called as order group) is broken down into several batches or picklists. Each batch/picklist is a collection of multiple order-lines belonging to several orders.  Each batch is assigned to a particular picker who then collects the picking tote/pick-vessel from the depot and travels to corresponding pick locations to complete all picks within the batch. The tote is then sent for further downstream consolidation process. Thus, the picker traverses a particular path within a pick-zone to fulfill a batch. The pictorial representation of Batch-Picking Process can be seen in figure 1.

Figure 1: Overview of Manual Batch Picking Process
Figure 1: Overview of Manual Batch Picking Process

The overall Batch-Picking problem is a two-part interconnected solution strategy. The first part is called order batching, where all the orders in the order group are assigned to a particular batch. The second part is called pick sequencing where the exact sequence of picks for each batch is decided. The overall objective of the problem is to minimize the total travel distance of all batches formed from an order group. Both Batching and pick sequencing are coupled with each other, meaning that only an optimal batching can result in optimal pick sequencing and pick sequencing in turn depends on the batching. If there are fewer orders and fewer pick locations in a warehouse, we can enumerate all the possible combinations and choose the best one. However, as the number of orders or pick locations in a warehouse increases, the total possible combinations increase exponentially making it almost impossible to enumerate all possible combinations in real time. This makes Batch-Picking a NP hard problem making it exceedingly difficult to solve optimally in a limited amount of run-time. 

Evolving challenges in Batch picking algorithms

Although a well-researched problem, there are several challenges in the implementation of Batch-Picking solutions at a customer site. The biggest challenge arises from the complicated process flow and multiple variants of Batch-Picking. Consider an online groceries delivery warehouse. In these types of warehouses, the customer orders typically consist of 15–20 line items. The number of unique items in the warehouse is limited thus increasing the commonality between the orders (there can be several orders which require 1kg of potatoes). In such scenarios, splitting the order across multiple batches followed by downstream consolidation is a very efficient picking strategy. This also requires the warehouse to use a single tote for collecting items. The weight of the SKUs (Stock Keeping Units) in this case is known accurately, however volume information is usually unavailable. On the other hand, if we look at a typical e-commerce warehouse such as Amazon, the average number of lines per item is typically 1-3, which is on the lower side, whereas the unique number of items offered is extremely high. In such scenarios, a cluster picking process is ideally implemented where the order is not split across batches and each order is picked into a separate compartment on the pick-vessel. Both weight and volume of items are known accurately and both constraints can be considered during batching. Since there can be several zones in warehouse, pick and pass process flow or even partial automation can be implemented.

These simple examples advocate the fact that each zone within a warehouse and each warehouse of different customers may implement/require vastly different variations of Batch-Picking solution. Customers might require only batching, batching & pick-sequencing or only pick-sequencing as a solution requirement. There can also be different flavors of picking depending on whether a single order can be split across multiple batches or not, the nature of pick-tote/vessel. The constraints on batch formation can be based on different combinations of total volume, total weight, number of all items and number of orders in a batch. Corresponding to all these various process flows, customer requirements, and their combinations, a total of 30+ different scenarios are possible. These different scenarios are summarized in figure 2.

Figure 2: Summary of 30+ Batch-Picking scenarios and workflows
Figure 2: Summary of 30+ Batch-Picking scenarios and workflows

There can be additional variations of process flows depending on how the pick tote is handled by the picker. In some cases, the picker must start from the depot and return to the depot to complete the batch. In other cases, the pick tote can be passed from one zone to another which is also called the pick and pass scenario. Further complications arise from the warehouse layout of each warehouse/zone. The number of aisles, cross-aisles, bays per aisle, positions per bay and even the location of the depot itself can vary for each zone/warehouse. Different travel restrictions can be imposed in warehouses to avoid congestion and collision of pickers. Another challenge with respect to Batch-Picking is that it is a continuously evolving process in warehouses domain. The most recent advents involve partial use of humans and partial use of AMRs (autonomous mobile robots) to complete the picks. Introduction of partial/complete automation of picking process can lead to even complicated process flows and numerous variations.

Current day batching algorithms in general are highly customized to the specific process flows/scenarios. However, with the number of scenarios and complexities of the pick process flows constantly growing with rapid evolution of e-commerce, batching algorithm ability to handle such diversity becomes paramount requirement. This is because the cost of modifications to a picking solution to customize at each customer/each customer site/each zone in a warehouse is significantly high and time-consuming and can become an impractical strategy. This raises an obvious need of Batch-Picking framework that has ability to be process adaptive, layout agnostic and robust so that it can be easily extended for diverse scenarios and easily be deployed at multiple customer sites with a mere change in configurations. At the same time, the solution must show a significant reduction in travel distance and improvement in picking throughput. The solution also must follow the latency requirements and generate an optimal batch picking strategy in a sufficiently small amount of run-time. Typically, less than 20% of average batch picking time is considered as upper cap for batch planning which ensures that the pickers are not idle due to delayed planning. These latency requirements are subject to change with the process workflow and warehouse layout itself and thus the solution must have the ability to adapt to different latency requirements. The solution can be integrated with WES (Warehouse Execution Systems) or can be provided as a standalone service to cater to a wide range of customers with minimal modifications to the algorithm itself. 

Solution Strategy

The Adaptive Algorithm for pick workflow agnostic Order Batching & Manual Picking is an intelligent order batching and pick-sequencing solution developed by Decision Intelligence (DI) team which is a part of Honeywell Intelligrated. This strategy was developed with an objective to encompass all these qualities that an ideal batching and pick-sequencing solution must have. This strategy is a combination of a deterministic approach and stochastic optimization approach. Deterministic approaches based on mathematical formulations ensure that we can find the find the optimal solutions, however the nature of this problem makes this strategy infeasible as the purely deterministic algorithm will need an extremely large run-time and computation power even for a moderate sized problem. On the other hand, evolutionary algorithms implement a stochastic optimization-based approach thus finding a near optimal solution within shorter run-time even with low computation power. Although this approach is useful for solving a lot of optimization problems, the Batch-Picking problem is difficult to model as a stochastic optimization problem and even a good model can only achieve slightly better results than the baseline. To address the shortcomings of both methods and exploit their advantages, we developed a clever strategy to decouple the batching and pick sequencing problems from each other and implemented a combination of deterministic and stochastic optimization approach to solve them respectively, independent of each other. The first phase of deterministic optimization creates optimal order batches without considering the exact travel distance and the second phase of stochastic optimization ensures that the pick sequence for each batch is optimal. This strategy ensures that we find a satisfactory solution which can achieve significantly high improvement in throughput even for a large problem in sufficiently small amount of run-time. Figure 3 gives an overview of the core working principle of this strategy.

Figure 3: Overview for Core working Principle of Smart Batching Strategy
Figure 3: Overview for Core working Principle of Smart Batching Strategy

The solution strategy has multiple configurations which can be set based on exact requirements and process flow of batch picking for each individual zone/warehouse. Thus, the one-time effort of this service development and WES integration will be able to cater to all possible Batch-Picking requirements. Even if further modifications are made in the process, they can easily be accommodated by changing the existing configuration. This will immensely cut down on the cost of development and customer specific customizations before each new deployment of the service. This strategy is also agnostic to the warehouse layout and location of depot within the warehouse. New evolutions in batch picking field including the partial/complete automation of process can be incorporated in existing solution owing to its unique way of breaking the problem into multiple sub-problems and solving them sequentially. Thus, this strategy not only supports the numerous existing process flows but can also cater to scenarios which may come up in upcoming years.

The strategy is extensively baselined for multiple scenarios and multiple sample problems with order group size varying between 30 to 100 whereas the pick locations in warehouse varying between 500 to 20000.  Although configurable, the algorithm run-time of 1.2 seconds per order was found to give optimal and consistent solutions. Through these experiments, the DI algorithm showed an average of 40% improvement in order picking throughput and 45% reduction in travel distance of the pickers. These results promise a significant boost in the efficiency of pickers, pushing the warehouse operating capacity to its limits. 

Our solution offers a strong decision intelligence engine which can optimally drive the picking process in any warehouse, irrespective of its layout and process workflow to boost the throughput and adapt to the changing needs and trends in e-commerce sector. Adaptation of such a robust and efficient strategy in warehouses will definitely be a game changer for years to come.

Check the Part 1 and Part 2 of this blog series.

To subscribe to the Honeywell Intelligrated blog and get exclusive communications and updates, click here.